Observability, Grafana, SCA

An early skeptic of DeepSeek’s (DS) claims of producing a $5.6 million blur-fast artificial intelligence (AI) model says the cost actually exceeded $1.5 billion.

In a report issued Friday, SemiAnalysis estimated the Chinese AI startup’s total server capital expenditure was a staggering $1.6 billion, with most of the spending to maintain and operate GPU clusters.

“Our analysis shows that the total server CapEx for DeepSeek is ~$1.6B, with a considerable cost of $944M associated with operating such clusters,” SemiAnalysis said in its report Friday.

The market researcher says the approximately $6 million training cost for DeepSeek V3 accounted for only GPU pre-training expenses and does not include research and development, infrastructure, and other crucial costs.

“We believe [DeepSeek] have access to around 50,000 Hopper GPUs, which is not the same as 50,000 H100, as some have claimed,” SemiAnalysis analysts said. “There are different variations of the H100 that Nvidia made in compliance to different regulations (H800, H20), with only the H20 being currently available to Chinese model providers today. Note that H800s have the same computational power as H100s, but lower network bandwidth.

“We believe DeepSeek has access to around 10,000 of these H800s and about 10,000 H100s. Furthermore they have orders for many more H20s, with Nvidia having produced over 1 million of the China-specific GPU in the last 9 months,” the analysis concluded.

The report did acknowledge that operational costs for DeepSeek could plunge five times by the end of 2025 because of DeepSeek’s ability to more quickly adapt compared to its larger rivals.

Another report, from The Futurum Group, casts similar doubt. “While the industry has been aggressively pushing to bring down the cost of training and democratize AI, the claims that a handful of NVIDIA H800s are doing this and the claims of this being genuinely open-source are hard to believe,” a team of Futurum analysts wrote. “There is a lack of transparency on what infrastructure was used and whether or not techniques beyond time scale were used.”

While the SemiAnalysis report agrees with DeepSeek that R1’s performance matched that of OpenAI o1 in reasoning, R1 was not the clear leader across all metrics. And while DeepSeek has gained attention for pricing and efficiency, the report said, Google’s Gemini Flash 2.0 is similarly capable – even cheaper – when accessed through API.

DeepSeek’s claims of its open-source model’s reasoning superiority were acceptable on one level, but what had the industry rocked and shocked was the alleged cost – a pittance compared to the billions of dollars that OpenAI, Microsoft Corp., Anthropic, Amazon.com Inc. and Alphabet Inc.’s Google are pouring into their model development. DeepSeek’s boasts led OpenAI, Microsoft and Meta Platforms Inc. to investigate the Chinese startup and its operational costs.

Executives from Microsoft ($80 billion) and Meta ($65 billion) said they intend to continue massive spending this year on AI.

“Three quarters of what DS is putting out there is smoke and the other 25% is mirrors, based on OpenAI’s research,” Danny Jenkins, co-founder and CEO of ThreatLocker Inc., said in an interview. “For Microsoft to create product is incredibly difficult. They are building something from the ground up. It is much cheaper to do something a second time (as DeepSeek did).”

HyperFRAME Research CEO Steven Dickens said while the analyst community speculates on where DeepSeek gleaned its GPUs and how many it had access to, “We see hyperscalers deploy the model, such as AWS in their Bedrock service, we start to get a perspective on the compute requirements, (and it is) safe to say the $6m is hyperbole, not reality,” he said.

TECHSTRONG TV

Click full-screen to enable volume control
Watch latest episodes and shows

Networking Field Day

TECHSTRONG AI PODCAST

SHARE THIS STORY